Vocabulary at the Living–Machine Interface: A Narrative Review of Shared Lexicon for Hybrid AI

Abstract

The rapid rise of bio-hybrid robots and hybrid human–AI systems has triggered an explosion of terminology that inhibits clarity and progress. To investigate how terms are defined, we conduct a narrative scoping review and concept analysis. We extract 60 verbatim definitions spanning engineering, human–computer interaction, human factors, biomimetics, philosophy, and policy. Entries are coded on three axes: agency locus (human, shared, machine), integration depth (loose, moderate, high), and normative valence (negative, neutral, positive), and then clustered. Four categories emerged from the analysis: (i) machine-led, low-integration architectures such as neuro-symbolic or “Hybrid-AI” models; (ii) shared, moderately integrated systems like mixed-initiative cobots; (iii) human-led, medium-coupling decision aids; and (iv) human-centric, low-integration frameworks that focus on user agency. Most definitions adopt a generally positive valence, suggesting a gap with risk-heavy popular narratives. We show that, for researchers investigating where living meets machine, terminological precision is more than semantics and it can shape design, accountability, and public trust. This narrative review contributes a comparative taxonomy and a shared lexicon for reporting hybrid systems. Researchers are encouraged to clarify which sense of Hybrid-AI is intended (algorithmic fusion vs. human–AI ensemble), to specify agency locus and integration depth, and to adopt measures consistent with these conceptualizations. Such practices can reduce construct confusion, enhance cross-study comparability, and align design, safety, and regulatory expectations across domains.

1. Introduction

Walk through any robotics or AI conference and you can overhear a babel of terms: cyborg tissues, joint cognitive systems, neuro-symbolic hybrids, human–AI teaming. Each phrase refers to some combination of the living and the machine, but no one fully agrees on how tight the coupling is, or where agency and accountability lie. This definitional inconsistency hampers theory-building and comparability across studies. The issue is especially troublesome for the biomimetics community whose creations, straddling the border between living and machine, demand precise language to guide design and communication.

Early nomenclature set the stage for the definitional confusion we see today. In 1960, “cyborg” was introduced as “the exogenously extended organizational complex functioning as an integrated homeostatic system unconsciously” [1], anchoring the term in physiological integration—an image familiar to humans. Three decades later, the concept was broadened, “A cyborg is a cybernetic organism, a hybrid of machine and organism, a creature of social reality as well as a creature of fiction” [2], shifting attention from hardware to sociopolitical identity.

Contemporary scholarship fragments the definitional landscape further. AI researchers and defense technologists narrow Hybrid-AI to the fusion of symbolic and sub-symbolic methods, emphasizing representational synergy [3]. Information-systems researchers prefer hybrid intelligence—“the complementary strengths of human intelligence and AI so that they can perform better together than either could separately” [4]. Cognitive engineering offers joint cognitive systems: two or more cognitive systems, of which at least one includes a human, collaborating to keep a complex process in control [5].

Viewed side-by-side, these labels converge on the idea of coupling the living and the machine, but they diverge on agency, depth of integration, and moral valence. For designers crafting surgical cobots or wearable exoskins, typical artifacts in biomimetics, such ambiguity frustrates literature reviews and hampers science communication. Furthermore, for theorists measuring trust or workload, such unclear concepts can threaten construct validity and cross-study comparability. This lack of a shared lexicon is particularly evident in the case of Hybrid-AI. Although Hybrid-AI has a precise technical meaning that refers to the integration of symbolic and sub-symbolic AI methods, it is often misinterpreted in interdisciplinary and public discourse as describing human–AI collaboration or hybrid intelligence. This semantic drift blurs the boundary between machine–machine architectures and human–machine systems, creating confusion across technical, social-science, and policy domains. As a result, scholars and practitioners frequently talk past one another, which hinders comparative research and the coordination of design and regulatory frameworks. Therefore, exploring the key terminologies emphasized across disciplines and organizing them into coherent categories is an essential step toward conceptual clarity.

While a number of recent reviews have examined related themes, most remain anchored within a single discipline or conceptual niche. For example, studies in production management focus on hybrid intelligence as an optimization strategy [6]. Research in marketing, consumer research, and psychology investigates AI’s influence on consumer behavior through bibliometric analyses of theories and themes spanning cognition, social media analytics, and technology adoption [7,8]. Literature review work in journalism traces the rise of automated and algorithmic reporting, using thematic analyses to explore how artificial intelligence is applied across news production and storytelling [9]. Other efforts center on individual constructs such as AI literacy [10], automation [11], augmented intelligence [12], or human-in-the-loop systems [13]. Together, these reviews provide valuable but fragmented insights confined to particular domains. The present review builds on them by integrating studies from multiple databases across engineering and social-science disciplines, offering a more comprehensive view of both human-centered and technical perspectives. Owing to these differences in scope and methodology, our analysis departs from earlier discipline-bound syntheses and contributes a cross-field lexicon that identifies converging research fronts, terminological fault lines, and directions for future inquiry in hybrid-AI scholarship.

To map the definitional landscape, we assemble a curated corpus of 60 hybrid machine system (HMS) definitions spanning engineering, HCI, philosophy, and policy. We use HMS as a collective term encompassing any arrangement in which biological or human agents and artificial systems jointly contribute to perception, reasoning, or control. Within that family, Hybrid-AI refers specifically to architectures or teams in which symbolic and sub-symbolic intelligence are integrated with human cognition. We (1) identify semantic clusters, (2) extract cross-cutting dimensions such as agency locus and integration depth, and (3) propose a faceted taxonomy portable across disciplinary borders. Clarifying where concepts overlap—and where they do not—gives biomimetics researchers, and the broader hybrid-systems community, a common lexicon for cumulative science and responsible innovation. At the same time, we show that the language of “hybridity” has intensified in the wake of large-scale AI and biohybrid breakthroughs since 2020, making the present moment an inflection point for terminology.

The paper proceeds as follows. Section 2 situates HMS terminology within earlier work on cybernetics, distributed cognition, and neuro-symbolic AI. Section 3 details our data-collection and coding procedures. Section 4 presents clustering results and semantic dimensions. Section 5 discusses our findings and the implications, while Section 6 outlines limitations and boundary considerations.

2. Literature Review

Scholarly discourse on HMS has followed the same line. Three partially insulated traditions—cybernetics and the cyborg imaginary, distributed or joint cognition, and the newer conversation on hybrid or augmented intelligence—have matured in parallel, each bringing its own priorities and concepts. Subsequent empirical work often inherits these framings wholesale, rarely pausing to notice the alternatives that neighboring fields provide. The result is conceptual drift.

The first lineage is physiological and control-theoretic. Mid-twentieth-century cybernetics treated the body as a platform to be re-engineered for hostile environments, introducing the cyborg concept as an organism whose homeostatic functions extend beyond the skin through engineered add-ons [1]. In the very same year, another leading computer scientist proposed “man-computer symbiosis,” predicting “very close coupling between the human and the electronic members of the partnership” [14]—a parallel vision that centered cognitive, rather than physiological, integration. Together these twin 1960 visions framed technology as both body and thinking partner.

That same logic now drives research on bio-hybrid actuators, tissue–electrode interfaces, and neural prostheses. Over time, cultural layers accreted: later scholarship reframed the cyborg as a site of identity and boundary politics [2]. Others extended the idea further, arguing that humans are “natural-born cyborgs” who habitually offload cognition onto notebooks, calculators, and—now—smartphones [15]. Such work foreshadowed today’s wearable or tissue-integrated sensors that blur organismal and artefactual boundaries [16]. Even within biomimetics, therefore, cyborgs can signal anatomical fusion or sociopolitical hybridity, depending on which branch of the lineage an author draws upon. Recent demonstrations, such as the creation of biohybrid cardiac tissue seeded with nanoelectronics [17,18] or light-guided “cyborg insects” for search-and-rescue [19]—show how the term continues to evolve in laboratory practice between bioengineering jargon and layperson-friendly metaphor.

A second tradition emerges from cognitive systems engineering and the anthropology of practice. Field studies of navigation crews shifted cognition beyond the skull, embedding mental work in networks of people, artifacts, and representations [20]. Building on that idea, the notion of a joint cognitive system casts humans and artifacts as co-controllers of safety-critical processes [5]. Here the analytic unit is the functional ensemble—checklists, displays, spoken commands, tacit know-how. Agency is explicitly distributed, and research gravitates toward breakdowns, handovers, and resilience rather than toward anatomical fusion or representational synergy. Recent human-autonomy teaming surveys reveal the tension, with authors finding that “construct confusion” across human-automation teaming, human–robot interaction, and joint-cognitive systems studies creates difficulty in accumulating a cohesive scientific knowledge base [21]. Similar calls for terminological clarity are a key motivator for our research here.

The third and most recent genealogy stems from artificial-intelligence research and human–AI teaming. Information-systems scholars describe hybrid intelligence as the complementary strengths of humans and AI combining to achieve complex goals while learning from one another [4]. Others extended this view, arguing that hybridity lies not in blending algorithms but in orchestrating complementary competencies [22]. Industry white papers rebrand the same idea as augmented intelligence, insisting that machines enhance rather than replace human workers [23]. Providing an explicit example, the American Medical Association (AMA) explicitly frames AI as augmented intelligence that should support, not supplant clinical judgment [24], highlighting the normative stakes of word choice.

This stream is optimization-minded. Hybridity becomes a design variable: allocate each sub-task to whichever agent, statistical or biological, performs best, then tune for accuracy, speed, or cost. Thus, in this stream, debates focus less on system boundaries and more on workflow orchestration, explainability, and real-time calibration of decision authority. Yet the technical community also advances a different sense of hybridity—“Hybrid-AI” as neuro-symbolic integration [3]—where the fusion occurs within the algorithm rather than between human and machine. Distinguishing these two senses is obviously important because conflating them obscures which research problems are architectural in nature versus those that tread into the sociotechnical realm.

All three lineages insist on non-trivial coupling between human and machine. All treat that coupling as instrumental. Where they part company is the locus of agency, depth of integration, and moral valence. Cybernetic framings see agency as an embodied extension; distributed-cognition work views it as emergent across socio-technical ecologies; hybrid-intelligence research treats agency as a dial set task by task. Integration ranges from prosthetic fusion through representational coordination to interface-level orchestration. Normative stances vary from evolutionary adaptation through resilience seeking to performance optimization, each embedding tacit assumptions about autonomy, liability, and trust. These patterned disagreements matter. A biomimetics study framed in cyborg vocabulary invites questions about tissue compatibility and metabolic load; one written in joint-cognition terms will be judged on coordination and error recovery; a manuscript positioned around hybrid intelligence must speak to learning curves and decision metrics. Moreover, popular fascination with “droids” and “humanoids” amplifies the stakes: media headlines about Tesla’s Optimus robot or “ocean-going cyborg jellyfish” preload public expectations with both awe and anxiety, making terminological precision central to science communication [25].

Terminology becomes even more tangled when AI enters the scene. A human-in-the-loop system may be labeled “augmented intelligence” in human–computer interaction (HCI), “joint cognitive system” in human factors, or simply “AI” in psychology. Such variation obscures meaningful distinctions between non-AI automation, pure AI, and Hybrid-AI—the term we foreground here. Although Hybrid-AI has a precise technical meaning—the integration of symbolic and sub-symbolic methods for reasoning and learning [3]—public and interdisciplinary discourse often conflates hybrid with human–AI teaming. The word evokes flesh-and-circuit synergy, not architectural fusion. Consequently, many healthcare, education, and urban-governance projects describe collaborative decision support as Hybrid-AI despite employing only neural models plus a human operator. Conceptual drift happens. Definitions matter.

This review therefore asks:

RQ1: How do disciplines define Hybrid-AI and related human–machine systems concepts?

RQ2: Which Hybrid-AI system types dominate current studies?

RQ3: Which assumptions still need unpacking to enable cross-disciplinary research?

Biomimetics: A Test-Bed for Terminological Collision

The study of biomimetics already showcases the complicated terminological map. A recent path-planning study combined multiple bio-inspired algorithms, dubbing the method a “hybrid swarm-intelligence algorithm” and reporting improved efficiency [26]. A 2025 investigation steered Endebius florensis beetles using backpack-mounted LEDs, labeling the creatures “cyborg insects” and “insect-machine hybrids” [19]. Earlier work wired microelectronics onto live jellyfish, producing a “biohybrid robotic jellyfish” that merged animal propulsion with on-board control [27,28]. Three papers, three labels—hybrid, cyborg, biohybrid (and even a mention of robots for good measure)—each pointing to living–machine fusion yet signaling different conceptualizations of it. The inconsistency underscores why a stable lexicon is urgent. Figure 1 organizes frequently conflated terms by theoretical focus (e.g., algorithmic architecture, human–AI interaction, and cognitive/distributed systems), clarifying where ‘Hybrid-AI’ (algorithmic) sits relative to ensemble notions such as hybrid intelligence or joint cognitive systems.

Recognizing the landscape, the remainder of this article treats genealogy as an analytic scaffold instead of historical anecdote. The contrasting assumptions about agency, integration, and normativity become coding dimensions applied to sixty verbatim definitions drawn from engineering, HCI, philosophy, and policy. The resulting comparative taxonomy aims to travel across disciplinary borders—and to give biomimetics researchers a clearer map for designing, evaluating, and naming the next generation of living-machine hybrids.

3. Methods

We undertook a narrative scoping review combined with a concept analysis to clarify how the terms AI, Hybrid-AI, and automation are defined and distinguished in human-centered contexts. Because the current debate is moving quickly, the search window was restricted to January 2020 through May 2025, and all database queries were executed in June 2025. Five bibliographic sources, PsycINFO, Scopus, the Communication & Mass-Media Complete index, IEEE Xplore, and LISA, returned ninety-one records. Citation chaining from key papers yielded thirty-four additional items, and manual checks contributed six more, giving a total of 131 hits. After removing ten duplicates and three records without minimally sufficient metadata, 118 items proceeded to full-text screening. Thirty-two were excluded for lacking definitional content or a human-facing relevance, leaving eighty-six reports that supplied sixty unique term-definition pairs.

The 2020–2025 corpus served primarily to identify high-frequency and widely discussed terms related to autonomy, AI, and Hybrid-AI, and to derive the analytic framework used for cross-disciplinary comparison. Focusing on this recent time period for term identification mitigated the risk of selecting older and/or out-of-use terms that would dilute our analysis. Because many papers in this corpus were exploratory and lacked explicit definitions, we subsequently conducted targeted follow-up searches without date limits to retrieve canonical or authoritative definitions of those terms. This two-stage strategy ensured both recency and conceptual completeness.

Eligibility hinged on five criteria applied in a single pass at full text: the work had to be peer-reviewed; it had to discuss definitions or conceptual boundaries of AI, Hybrid-AI, or automation; it had to address human use or impact rather than purely technical performance; it had to fall within the disciplines of human–computer interaction, human factors, psychology, communication, information or computer science (with occasional adjacent inclusions where definitional work was substantive); and it had to be published in English. Applying these rules produced a tractable but cross-disciplinary corpus that spans laboratory studies of joint cognitive systems, qualitative analyses of human–AI teaming, and technical proof-of-concepts for neuro-symbolic architectures used in healthcare or education.

Table 1. Coding Operational Sheet for Classifying HMS-related Concept.

Dimension	Code	Definition	Decision Criteria
Agency Locus	(H) Human-dominant	Human retains primary decision-making and control	Final approval, veto rights, or manual takeover is explicit.
	(HM) Shared/Dynamic	Control dynamically allocated or jointly shared between human and machine	Requires all three conditions: (a) Human execution channel explicit; (b) Machine execution explicit; (c) Shared control/handoff mechanism present.
	(M) Machine-dominant	Machine retains primary control; human mainly sets goals or monitors	Machine autonomy or automatic policy execution stated; no human override mentioned.
Integration Depth	(L) Loose coupling	Functional complementarity, low interdependence; systems run in parallel	No human actor mentioned in the definition or only for advisory/approval/offline interaction; human & machine loosely connected or separate.
	(M) Moderate coupling	Substantial information exchange and feedback loops without physiological embedding	Human actor is explicit, and coordination/feedback/interaction cues are present, but no physiological/biomechanical integration.
	(H) High coupling	Physiological, cognitive, or mechanical embedding and sustained, continuous interaction; human and machine form a unified operational system	Human and machine operate as a unified operational system; explicit reference to body/device-level integration & closed-loop interaction.
Normative Orientation	(1) Positive stance	Highlights benefits, efficiency, capability enhancement, and empowerment	Assumes trustworthiness, improved autonomy, or better outcomes.
	(−1) Negative stance	Highlights risks, harms, threats, or ethical concerns	Assumes loss of control, liability concerns, or erosion of trust.
	(0) Neutral stance	Purely descriptive or definitional, no explicit value judgment	Focuses on what it is, not what it does.
	(1,−1) Mixed stance	Contains both positive and negative normative signals	Reflects uncertain trade-offs, dual-use nature, or socio-ethical complexity.

details the coding rubric and decision rules. Notably, shared/dynamic agency (HM) required three explicit conditions (human execution channel, machine execution channel, and an articulated hand-off).

The unit of analysis for data extraction was the concept entry: a term, its verbatim definition, and its bibliographic source. Each concept was scored on three axes. We examined every definition to identify explicit or implicit cues for the three dimensions. The detailed coding scheme, including operational definitions and decision criteria, is summarized in Table 1. Agency locus captured where primary task initiative and authority reside and was encoded as human-led 0, shared 1, or machine-led = 2. Integration depth measured how tightly human and machine elements are coupled, ranging from loose 0, through moderate 1, to high 2. Normativity registered the overall valence of the framing, with positive +1, neutral 0, and negative −1. This coding exposed patterned disagreements that anecdotal reading tends to miss. For clustering, we collapsed −1/0/+1 to 0/1 to balance classes; the three-level distributions are reported descriptively. All entries were consolidated in a master sheet and coded numerically in a companion sheet (combined in Appendix A.1).

We also consulted contextual information when wording was ambiguous (e.g., the disciplinary domain of the source, such as HCI vs. control systems, stated assumptions about autonomy, or typical use cases) to ground coding decisions and avoid over-interpretation. Coders recorded brief evidence notes (e.g., keywords like approval/override, shared autonomy/teaming, autonomous/automatic) to document the rationale for each assignment. Two coders independently coded all definitions using the shared codebook. Inter-coder agreement (percent agreement) was high across the three dimensions: 91.7% (agency locus), 90.0% (integration depth), and 86.7% (normative orientation). Discrepancies were resolved through discussion to consensus, with the agreed labels constituting the final dataset for analysis.

To illuminate higher-order structure in the dataset, we submitted the three numerical codes to a k-means cluster analysis after standardizing the scale. Optimal k was determined by inspecting k = 3–5 and then maximizing silhouette width and retaining theoretical interpretability. Cluster selection was guided by two criteria: (1) the silhouette score, higher than 0.5 indicating moderate-to-good separation between clusters, and (2) a minimum cluster size threshold of 10 concepts to ensure interpretability and conceptual stability. Among all tested solutions, the k = 4 configuration maximized the silhouette coefficient (0.565) while meeting the size constraint, providing the most theoretically grounded and statistically robust solution.

The resulting clusters map four regimes of hybridity. One is machine-led and loosely integrated, another is shared-agency with moderate-to-high integration and an optimistic tone, a third is human-led with medium coupling and a neutral stance, and the fourth is low-integration, human-centric, and positive. Cluster sizes were nineteen, fifteen, fifteen, and eleven, respectively, and centroids for agency, integration, and normativity are reported later in

Table 2. Cluster Sizes & Composition.

Cluster	Label	Size (n)	Percentage
C0	Machine-Led Low Integration	15	25.00%
C1	Shared Collaborative Normative	19	31.70%
C2	Human-Led Medium Integration	15	25.00%
C3	Human-Centric Low Integration	11	18.30%
Total		60	100%

. Figure 2 visualizes the integration–agency landscape colored by these four clusters; the full coding rubric and membership roster appear in the Appendix A.

Table 3. Distribution across agency, integration, and normativity (n = 60).

Dimension	Level/Value	n	% of Total
Agency	Human-led	21	35.0%
	Shared	25	41.7%
	Machine-led	14	23.3%
Integration	Low	24	40.0%
	Medium	29	48.3%
	High	7	11.7%
Normativity	Negative	1	1.7%
	Neutral	36	60.0%
	Positive	23	38.3%
Means	Agency = 0.88  Integration = 0.72  Normativity = 0.53

summarizes the distribution of Agency (H/HM/M), Integration (L/M/H), and Normativity (−1/0/+1) across the 60 concepts.

Taken together, the review supplies a disciplined vocabulary map rather than another list of buzzwords. By grounding definitions in explicit agency, coupling, and normative coordinates, the taxonomy can travel across disciplinary borders and—of special interest to Biomimetics readers, help designers of bio-hybrid actuators, cyborg insects, or hybrid swarm-intelligence algorithms decide what kind of “hybrid” they are really building.

The five-year window brackets what is arguably the fastest vocabulary acceleration in AI history, driven of course by the explosion of AI discourse after ChatGPT’s public release in December 2022 and preceding discussions around events like the introduction of AI-specific regulation in several nations. Restricting the search to 2020–2025 therefore captures the inflection point at which hybrid terminology migrated from specialist workshops into mainstream policy and press coverage. It also guards against retrofitted definitions applied to technologies that pre-date the current surge, ensuring the corpus reflects the language environment that current researchers and their lay audiences actually inhabit.

4. Results

The search conducted in June 2025 returned 131 records across five databases, citation chaining, and manual sources. After removing ten duplicates and three items with insufficient metadata, 118 records were assessed at full text. Thirty-two failed eligibility–typically because they offered no explicit definition or addressed purely technical performance, leaving 86 studies that anchored 60 distinct term-definition entries.

4.1. Concept Characteristics Before Clustering

Across the 60 entries, agency was most often described as shared between human and machine (41.7%), followed by human-led arrangements (35.0%) and a smaller contingent of machine-led systems (23.3%). Integration depth clustered around the middle of the scale: almost half of the concepts (48.3%) specified moderate coupling, 40.0% referred to loose tool-like links, and only 11.7% invoked high degrees of fusion. Framing was overwhelmingly neutral to optimistic—60.0% of definitions adopted an ambivalent or conditional tone, 38.3% were explicitly positive, and a single entry portrayed the technology in a negative light. Mean numerical scores across the three axes were 0.88 for agency, 0.72 for integration, and 0.53 for normativity, signaling a slight tilt toward shared control, moderate coupling, and favorable rhetoric (Table 3).

4.2. Cluster Landscape

Applying k-means to the three numeric axes yielded a four-group solution with a silhouette coefficient of 0.565. The clusters capture distinct regimes of hybridity. Machine-Led Low Integration Dominance (n = 15) aggregates concepts such as Hybrid-AI, neuro-symbolic architectures, and mixture-of-experts language models. These definitions emphasize algorithmic autonomy while keeping the human at arm’s length and adopt a largely neutral stance. Shared Collaborative Normative (n = 19) gathers terms like hybrid intelligence, mixed-initiative systems, and coactive design; here agency is balanced, integration ranges from moderate to high, and the tone is clearly upbeat. Human-Led Medium Integration (n = 15) encompasses AI-in-the-loop, joint cognitive systems, and decision-support tools, describing scenarios where humans retain primacy yet interact continuously with AI. Finally, Human-Centric Low Integration (n = 11) houses participatory AI, interactive machine learning, and RLHF, all framed as empowering technologies with light coupling and an explicitly positive ethos. Table 2 shows that Shared Collaborative Normative is the largest, followed by MachineLed Low Integration, HumanLed Medium, and HumanCentric Low.

Table 4. Cluster Centroids & Top Concept.

Cluster	Label	Agency Centroid	Integration Centroid	Normativity Centroid	Top Concepts
Cluster 0	Machine-Led Low Integration	1.87	0.00	0.13	Hybrid-AI, Neuro-Symbolic AI, Multi-Agent System
Cluster 1	Shared Collaborative Normative	1.00	1.26	1.00	Hybrid Intelligence, Adjustable Autonomy, Shared Autonomy
Cluster 2	Human-Led Medium Integration	0.40	0.93	0.00	AI-in-the-Loop (AITL), Immersive Technology (XR), Virtual Reality (VR)
Cluster 3	Human-Centric Low Integration	0.00	0.45	1.00	Participatory AI, Augmented Decision Making

reports the centroid values for the three key dimensions across the four clusters. The listed top concepts illustrate how each cluster captures a distinct locus of meaning within the hybrid human–machine systems (HMS) landscape. Hybrid AI and Multi-Agent Systems in Cluster 0 embody the machine-led, low-integration end of the spectrum, where control and decision-making are primarily algorithmic and human input is minimal. Hybrid Intelligence and Shared Autonomy in Cluster 1 represent shared-control paradigms that emphasize mutual adaptation and complementary capabilities between humans and AI. AI-in-the-Loop and immersive technologies such as XR/VR in Cluster 2 mark a transition toward human-led yet technologically embedded arrangements, where interaction is continuous but the human remains central to control. Finally, Participatory AI and Augmented Decision Making in Cluster 3 anchor the human-centric pole, emphasizing inclusion, transparency, and empowerment in socio-technical systems. Figure 2 positions every concept along two coding dimensions: Agency Locus (x-axis) and Integration Depth (y-axis), with color indicating Normative Orientation (neutral vs. valenced). Cluster polygons outline the semantic regions identified by k-means analysis, and stars mark the centroids.

Cluster compositions reinforce these qualitative signatures. Eighty-seven percent of C0 entries assign primary control to the machine, whereas C3 is entirely human-centric. C1 alone contains high-integration definitions (26.3% of its items), echoing claims of deep symbiosis. Normative orientation also sorts cleanly: the two clusters dominated by human agency show the strongest positive framing, while neutral language prevails in C0 and C2. The plot reveals four distinct groupings: machine-led, loosely coupled systems (C0); shared-control, moderately to highly integrated systems (C1); interactive, human-led frameworks with moderate coupling (C2); and human-centric, low-integration models (C3)—together forming a continuum from technical autonomy to human-centered collaboration.

A working glossary of high-salience terms with their coded attributes and the full 60-item list is provided in the Appendix together with cluster assignments. CSV files underpinning all tables and figures accompany the data-availability statement.

4.3. Robustness Check

Alternative k values between three and six produced lower silhouette scores and less interpretable partitions; the four-cluster solution therefore represents both the best statistical fit and the clearest conceptual map.

5. Discussion

Our findings shed light on how different disciplines articulate the idea of “hybridity” at the living–machine interface. The literature reveals some consistent yet divergent interpretations that structure the definitional landscape. Answering our first research question, across computer science, HCI, biomimetics, psychology, communication, and information science, we observe two dominant senses of “hybrid”: (i) algorithmic fusion (e.g., Hybrid-AI, neuro-symbolic models) that is typically machine-led with loose coupling, and (ii) human–AI ensembles (e.g., hybrid intelligence, mixed-initiative systems) that emphasize shared or human-led agency with moderate coupling. Investigating our second research question (which types dominate), we find shared or dynamic arrangements are most common, while high-integration concepts remain rare. Our third research question (assumptions to unpack) reveals the complexity of the landscape because “Hybrid-AI” is applied to both algorithmic fusion and human–AI teaming.

Terminological ambiguity carries practical risks that extend beyond academic debate. When the same term is used inconsistently across disciplines, policymakers may ground regulations in divergent meanings, producing gaps in accountability or over-regulation of systems that differ in intent and architecture. In media coverage, such ambiguity could lead to public misperception, framing Hybrid-AI either as a threat to human autonomy or as a promise of effortless collaboration. Both frames are a distortion of how these systems actually function.

The present definitional landscape presents many discussion points, one of which is the valence asymmetry in our data. Of the sixty concept entries we analyzed, all but one are framed either neutrally or optimistically. In other words, 98 percent of the literature portrays hybrid systems as neutral, benign, promising technologies. This positive frame contrasts sharply with risk-oriented narratives that dominate popular media and policy debate, where job displacement, bias, and existential threat often take center stage. The gap is consequential, beyond venue effects, it can shape risk registers, regulatory scoping, and public expectations. Because our corpus draws primarily from design-forward, solution-oriented venues, the imbalance is understandable, but it still matters. When researchers speak the language of benefit while the public hears the language of harm, expectations (and even prospective regulation) can drift out of sync between scientists and society. Bridging that gap suggests small shifts in research practice, e.g., explicitly acknowledging foreseeable failure modes and normative assumptions in definitional work. Doing so keeps the scientific lexicon aligned with the realities of governance and public trust without dulling the optimism that motivates innovation.

Perhaps related to this valence pattern is the most consequential ambiguity in our data: two radically different senses of “Hybrid-AI.” On one side sit machine-meets-machine architectures such as neuro-symbolic pipelines, mixture-of-experts language models, hybrid swarm algorithms, etc. These entries gather in Cluster 0, characterized by machine-led agency and a general tone that is more technical and less moral or ethical. On the other side sits man-meets-machine teaming. Examples include mixed-initiative systems, coactive designs, and participatory learning loops. These populate Clusters 1 through 3 and carry the social and ethical burden of hybrid designs. An implication for researchers is that the risk of placing both families under the same label could obscure design requirements or evaluation criteria, and perhaps even the anticipated regulatory/accountability framework the technology will operate in. An algorithmic-fusion system typically (in some regulatory systems) may fall under software-safety regulation, whereas a human–AI ensemble may follow workplace or medical-device statutes. Label clarity therefore guides not only experimental design but also compliance pathways and insurance underwriting [24]. Therefore, the point is emphasized again that label clarity, while seemingly pedantic at times, smooths the development of new technologies in addition to aiding cumulative science.

Our single negative-valence concept in the entire dataset (algorithm aversion) is notable for being a behavioral construct rather than an architectural label. Algorithm aversion describes humans’ avoidance of deferring decisions to machine judgment (e.g., decision aids, autonomous agents, etc.). Negativity thus centers on adoption psychology rather than hybrid engineering per se. The implication is that design efforts aimed at transparency and trust repair (following errors) are important, even when designers are highly confident that systems will perform well. Some trust research show that explanations can flip aversion into appreciation within several interaction rounds [29,30], suggesting that the attitude effects are changeable rather than fixed.

Our results also show that the promise of true fusion—living and machine united into a seamless whole—remains more rhetoric than reality. Only seven of the 60 concepts register high integration. Most entries settle for a framing that suggests only moderate to loose coupling, even when framing elsewhere in the paper points to deeper integration. Thus, two futures are visible. One, a convergence trajectory, would see advances in AI and biomaterials pull today’s moderate designs toward tighter living-machine fusion. The other, an asymptote trajectory, would keep integration bounded by current constraints, consolidating shared-control systems rather than dissolving the boundary. The data signal that, for now, the field leans toward the latter. Tracking this trajectory is a key goal for future research investigating the ever-evolving conceptualizations of hybrid technologies.

Finally, the public imagination remains a wildcard. Sensational headlines about “cyborg soldiers” or “humanoid butlers” can overshadow careful laboratory language [25]. Because biomimetic prototypes often feature living material, mislabeling can provoke bio-ethical backlash—even when the work is benign. Clear messaging can pre-empt misunderstanding and foster informed debate, ensuring that innovation and societal trust grow in tandem.

Where, then, do the terms most familiar to Biomimetics readers reside on this map? Generally, most cluster in human-led or shared-agency regimes with moderate integration and neutral-to-positive framing. That placement aligns with the community’s historical emphasis on safety. It also implies a practical editorial guideline: when authors invoke “Hybrid-AI,” they should specify whether they mean algorithmic fusion or human–AI teaming, and they should state both the locus of agency and the depth of integration up front. Doing so will let reviewers and fellow researchers evaluate the newest technologies against the appropriate yardsticks.

Several limitations frame these findings. First, the scoping approach deliberately privileged human-centered, definitional work. Risk-heavy or policy-oriented studies are under-represented. Second, the three axes—agency, integration, normativity—compress nuance; shared autonomy, for example, covers a spectrum of task-allocation schemes that our coding treats as equivalent. Third, clustering relied on a binary normativity code to preserve statistical balance, whereas descriptive statistics retained the full three-level scale. Fourth, the four-cluster solution balanced silhouette fit with interpretability; finer subtypes undoubtedly exist. The taxonomy explains how concepts are used, not how systems perform; linking cluster membership to safety, workload, or error-recovery metrics remains future work. Finally, we searched and analyzed only English language sources and definitions.

6. Conclusions

Our comparative taxonomy offers a lexicon for reporting hybrid systems. Authors should consider stating which sense of Hybrid-AI is intended (algorithmic fusion vs. human–AI ensemble), specifying agency locus and integration depth, and choosing yardsticks that match the conceptualization. Doing so reduces construct confusion and improves cross-study comparability. More broadly, such practices may better align design, safety, and regulatory expectations across domains. As our research is a building step, we encourage future research on this topic so as to maintain an open, living glossary that records ongoing changes as the area evolves.

Taken together, the evidence suggests that the center of gravity in hybrid research is not full biological-machine fusion but competent, shared control. Algorithmic fusion—the technical heart of Hybrid-AI—remains largely value-silent, while ensemble practices bear the ethical weight. The practical ask for Biomimetics authors is simple: state which sense of “Hybrid-AI” you mean, and specify agency and integration explicitly. Precision in language is the first step toward designs that are safer, more comparable, and ultimately more useful across disciplines.